What 2007 means to your data center
The SMP revival
Part I: It's fashionable these days for vendors to boast about blade servers and clusters of thin kit, but a large, multiprocessor server revival is underway - a trend that could have a massive impact on data centers.
Swift one- and two-processor servers have stolen some of the glory once owned by SMP (symmetric multiprocessing) systems. Faster processors and tight budgets have pushed customers toward using slim kit to handle business software and scientific workloads of all shapes and sizes. You've all heard the buzz, stirred up most by Dell, that Windows and Linux two-ways have guaranteed the demise of Unix giants.
But the differences between low-end and high-end servers and blades and big boxes are about to become more blurred than Rush Limbaugh's vision after a morning OxyContin. In 2007, all of the major server chip makers - Intel, IBM and Sun Microsystems - will foist new brands of multicore processors on the market. These chips with anywhere from four to eight cores will create a type of SMP-on-a-chip - a move that turns a one-way server into an incredible workhorse.
This multicore technology isn't radical in the sense that processor cores are constantly getting faster and smaller. More horsepower in a smaller space is a trend the computing industry knows well. This progression, however, speeds up in 2007, creating a dramatic shift that could see customers rethinking their data center plans.
"This trend has always been happening, but the rate at which it happens will accelerate over the next few years," said Jason Waxman, director of multiprocessor platform marketing at Intel. "It is certainly going to be a nice jump and boost for the IT industry."
All of the major vendors agree that there will still be a place for blade servers and standard one- and two-way systems. The great variety of software jobs requires that a great variety of servers exist.
That said, the leap from the dual-core chips on the market now and arriving in 2005 to the more sophisticated multicore chips that will be available in 2007 will change the server game. With the help of software advances, customers will be able to carve up these SMPs-on-a-chip into multiple blade-like systems or run the boxes as one, large computer.
"The first thing that hits all of us, when you think about all this performance, is that a four-processor system in 2007 will deliver the type of performance that a 32-processor or 64-processor system can produce today," Waxman said.
The most immediate gain for customers from this trend is the performance they will get per square foot of data center space. Even with a significant increase in compute demand over the next three years, the average customer could expect a serious decrease in data center size.
In addition, a new breed of customers should gain access to the kind of horsepower reserved for big spenders today. These multicore processor-based systems will require loads of memory to keep the chips happy, but even with that cost, vendors expect the overall price of a decent-sized SMP to come down.
All you can eat
Looking at Intel, for example, customers should note that the company expects Itanium and Xeon chips to be at a price parity by 2007. Intel's Itanium line is also ahead of the Xeon line in multicore technology.* This means that current Xeon customers could step up to higher-performing Itanium kit without having a sticker shock spasm.
Sun has a similar story but adds its own unique twist.
Sun recently turned to Fujitsu for SPARC help, signing on to use the SPARC64 processor instead of UltraSPARC from 2006 on. It's Fujitsu's product that will more or less go head-to-head against Intel's Tukwila version of Itanium and IBM's Power6 processor in 2007. All of these vendors are a bit cagey on exactly how many cores their chips will have at that time, but the best bets point to four cores each.
With "traditional" processor costs offloaded, Sun is looking to help customers out with more aggressive multicore designs in the form of its Niagara and Rock processors. The Niagara chips are due in early 2006 with 8 processor cores each, the Rock products should arrive by 2007 with fewer but more powerful cores. In both cases, Sun plans to throw lots of cheaper, slower processors at software instead of a more expensive, faster processor along the lines of SPARC64.
Sun expects the Niagara and Rock processors to outperform its current UltraSPARC III chips by 15x and 30x, respectively. These gains, however, mostly come with multithreaded software that lends itself to being spread across numerous cores. We'll have more next week on the software part of the multicore processor equation in Part II of this story.
Sun sees the average cost of an SMP system coming down as a result of this technology; but, of course, expects to sell more systems as customers make use of the power boost.
"You can imagine taking a large machine that can handle 72 software threads today, multiplying that by 10 and offering it in a single box to compute hungry customers," said Marc Tremblay**, a Sun Fellow and chief architect. "Or you can see a new class of customers getting a four-chip Niagara system that is essentially a 32-way SMP at a low cost."
"We are convinced that there is a huge appetite for compute cycles out there."
IBM was the first major vendor to roll out a dual-core high-end chip in the form of Power4 and has gone with a dual-core design again with Power5. The company has been very quite about what Power6 might look like, but you can see IBM making similar moves to its rivals.
With Power5, IBM benefitted from an improved manufacturing process that should make low-end and midrange servers based on the chip more affordable. In addition, IBM has managed to put an offshoot of Power - the PowerPC 970 or Apple's G5 chip - into low-end blade servers. By 2007, IBM should push these trends with Power6, although there is some speculation from industry insiders that IBM will have much less radical designs than its rivals. Although, funny as it sounds, the "Cell" chip being designed by IBM, Sony and Toshiba for entertainment consoles looks interesting.
Size kind of matters
As you'll see in the software part of this story, some vendors expect to see more dramatic gains out of these mini-SMP systems than others. There are a host of issues around partitioning, management and software licensing that control just how profound of a leap multicore processors may be.
Itanium user HP, for example, appreciates some of the benefits coming with the new chip technology but does not think a massive rebirth of the SMP is under way.
"There are some workloads out there that like and will continue to like to have a really powerful processor core for each thread," said Brian Cox, worldwide product manager for HP servers. "If you start stacking too many processor cores going through single sockets, choke points could pop up.
"So, there is going to be room for single core and dual-core chips for quite some time. We have to be careful about (multicore designs), as vendors, to make sure it's not technology for technology's sake but that this will benefit the customers."
It's no surprise to see chip-makers such as Intel and Sun be a bit more bullish about the technology. If nothing else, the engineers at these respective companies are excited to try their hands at something new, and it shows.
But, while radical multicore designs might not cause a revolution, they will create a vast new set of choices for customers.
Customers should more or less be able to "scale-out" within a single system. Instead of stacking 8 blade servers in a rack, a customer could consider each processor core as a single blade. Those willing to pick the multicore route over racks of thin servers could expect to have serious management and data center size benefits. Forget server sprawl when a a 42U rack can hold 300 64-bit processor cores.
And, as the software story will show, vendors at ease in the SMP world - Sun, HP and IBM - could have massive advantages over a vendor such as Dell in this model. Most of the tools and software that take advantage of the SMP-on-a-chip concept are tied to various versions of Unix. But by 2007 that could well change.
All of this means that shorter, squatter servers based on multicore chips will obliterate the line between scaling out and scaling up. Does this mean "shared everything" boxes will be hip again? It looks that way. ®
*It's hard to say exactly how this will play out given the recent changes to Intel's Xeon roadmap, but at present dual-core and multicore Itaniums are set to arrive ahead of their Xeon counterparts
**Are chip engineers supposed to look like this?